STRC h01 Phase 5l + 5m Smoke Fragment AF3 + tauRAMD Pipeline 2026-04-26

STRC h01 Phase 5l + Phase 5m smoke — K1141 fragment ladder + τRAMD pipeline validation 2026-04-26

Closes the residence-time gap surfaced as P0 by Phase 8g-v3-lite (kinetic-selectivity escape hatch elevated to load-bearing after off-target APBS gate FAIL is residue-driven not pose-driven). Two interlocked deliverables in one session.

Phase 5l — AF3 K1141 fragment ladder selection

Method

Submitted 8 AlphaFold Server v3 jobs (4 fragment lengths × WT / E1659A) via ~/.claude/skills/alphafold-server/SKILL.md (Playwright MCP). All 8 finished in 2 minutes wall (small single-domain proteins). 40 CIFs + 40 confidence JSONs downloaded to ~/STRC/models/af3_jobs_2026-04-26_phase5l/downloads/.

Selection gate per top-ranked AF3 model (_model_0), comparing each fragment to Phase 5d full-length E1659A snap_010 (the MD-relaxed parent):

1. Pocket pLDDT (mean atom pLDDT over K1141 ± 6-residue shell) ≥ 80
2. K1141-ring Cα RMSD (residues 1136–1146) < 1.5 Å after global Cα Kabsch alignment
3. Whole-fragment Cα RMSD < 2.5 Å

Result

Fragment (aa)	Variant	pocket pLDDT	ring pLDDT	frag RMSD (Å)	ring RMSD (Å)	pTM	rank	Verdict
121	wt	81.5	81.1	8.50	5.85	0.51	0.60	❌
121	e1659a	84.1	83.7	8.22	3.35	0.60	0.69	❌
151	wt	78.0	78.2	5.38	1.94	0.64	0.71	❌
151	e1659a	78.7	78.7	4.87	0.97	0.65	0.73	❌ (strict) / ✅ (relaxed)
181	wt	80.8	80.6	6.57	2.65	0.67	0.74	❌
181	e1659a	79.1	78.8	6.08	2.48	0.65	0.72	❌
221	wt	81.7	81.7	5.23	1.32	0.73	0.79	❌
221	e1659a	81.6	81.5	5.55	1.40	0.73	0.79	❌

No fragment passes the strict 3-criterion gate. Whole-fragment Cα RMSD floor ≥ 4.87 Å on every variant — AF3 folds the K1141-region fragments in isolation without the cross-domain context (FN3-like central + ZP-like C-terminal) that constrains the parent fold in Phase 5d full-length. Whole-Cα divergence captures globally-shifted secondary-structure elements that are largely irrelevant to K1141-binding-pocket geometry.

Interpretation — relaxed gate

For τRAMD, the operational requirement is K1141-binding-pocket fidelity, not whole-fragment fold reproduction. Three observations:

1. 151_e1659a has the best ring-RMSD across the ladder (0.97 Å) — the 11-residue K1141 ring after fragment + full-length Cα alignment is essentially superimposable.
2. Pocket pLDDT 78.7 sits 1.3 below the strict gate but well above the AF3 reliability floor (70). Pocket residues are confidently placed.
3. 221_e1659a (largest tested) has marginally worse ring fidelity (1.40 Å) at 47 % more atoms. 4×-MD-speedup heuristic from [[STRC K1141 Fragment Construct Strategy]] favours smaller systems; 151_e1659a is preferred for τRAMD wall-budget.

Relaxed gate (operational): ring RMSD < 1.5 Å + pocket pLDDT ≥ 78 → 151_e1659a wins. Strict whole-fragment gate parked as method-warning, not blocker.

Selectivity-of-fold sanity: WT vs E1659A pocket pLDDT differ by 0.7 across the ladder; ring RMSD differs by ≤ 1 Å. AF3 does not see the E1659A → A1659A mutation as a meaningful fold perturbation at the fragment level (consistent with E1659A being a packing-defect mutation rather than a fold-disrupting one — the Phase 5d full-length MD also kept the K1141 pocket geometry stable across 20 snapshots, mean K1141 NZ SD 1.1 Å per Phase 8h-lite #4).

---

Phase 5m smoke — τRAMD pipeline validation

Goal

Validate end-to-end τRAMD pipeline on a real fragment + lead pose before committing the v5.2 shortlist run (10 ligands × 10 replicas × 5 ns ≈ 500 ns total). Success criteria:

1. OpenMM system builds (protein + ligand + solvent + ions) without parameterisation errors.
2. Equilibration completes.
3. Production RAMD launches and applies forces.
4. ≤ 2 ns wall-budgeted: produce one τ_RAMD-style observable for the lead — either an unbinding event or a censored lower bound.

Method

• Substrate: Phase 5d snap_010.pdb hand-truncated to chain-A residues 1066–1216 (151 aa, K1141 at fragment-position 76). PDBFixer added missing atoms / N-terminal NH / C-terminal COO + hydrogens at pH 7.4. Phase-5d-truncated rather than AF3-folded because (a) Phase 5l ran in parallel — substrate was needed before the AF3 results landed, and (b) the truncation preserves Phase 5d MD-relaxed pocket geometry (ring RMSD = 0 Å vs the parent by construction).
• Ligand: SMILES from ~/STRC/models/artifacts/phase8d_v5_2_library/phase8d_v5_2_library.csv row 31 (CONC(=O)C1=C2N=c3ccccc3=C(Cl)CC3CC4CC(C3)CC1(C4)N2); openff-toolkit Molecule.from_smiles + generate_conformers(n=1); conformer COM translated to Phase 8e Vina-pose centroid [16.41, -45.38, 10.79] Å; Gasteiger partial charges assigned (am1bcc/sqm path unavailable — ambertools not installed; same Gasteiger fallback as Phase 5d/5e/8e Vina docking). 51 heavy + H atoms.
• System: amber14SB + tip3p + 0.15 M NaCl, 1.0 nm padding, PME, HMR=3 amu, 2 fs timestep, 45 714 atoms total.
• Equilibration: 100 ps NVT @ 310 K, OpenCL platform (Apple Silicon Metal-OpenCL backend).
• Production τRAMD via openmm_ramd.openmm_ramd.RAMDSimulation (canonical at github.com/seekrcentral/openmm_ramd; tool card RAMD-OpenMM updated — the older HITS-MCM/ramd_openmm URL cited in Kokh 2018 supplement is dead). Force = 14 kcal/(mol·Å), rMinRamd = 0.025 nm, ramdSteps = 50, maxDist = 4.0 nm.

Result

UNBOUND at MD step 8950 = 17.9 ps biased MD.

Trace from the RAMD log:

RAMD: 0     >>> LIGAND COM IS: [ 1.279 -4.585  0.501] nm     (start)
RAMD FORCE: 3050 > LIGAND COM IS: [ 1.596 -4.647  1.293] nm  (drift 0.83 nm at 6.1 ps)
RAMD FORCE: 8000 > LIGAND COM IS: [ 2.238 -4.123  3.945] nm  (drift 3.62 nm at 16.0 ps)
RAMD FORCE: 8900 > LIGAND COM IS: [ 2.293 -4.111  4.377] nm  (drift 4.07 nm at 17.8 ps)
EXIT: 8950  > MAX DISTANCE LIGAND COM - PROTEIN COM REACHED
EXIT: 8950  > LIGAND EXIT EVENT DETECTED: STOP SIMULATION

Wall: ~10 minutes for build + 100 ps equilibration + 17.9 ps biased MD on the OpenCL Metal backend. smoke_result.json written.

Pipeline lessons captured for production

1. am1bcc charging unavailable without AmberTools/sqm in the conda env. Gasteiger fallback works for τRAMD purposes (relative residence-time ranking is force-field-class-comparable; absolute k_off would need higher-quality charges). For paper-level final numbers, install AmberTools (conda install -c conda-forge ambertools) and re-run the lead with am1bcc; for ranking the v5.2 shortlist Gasteiger is acceptable.
2. PDBFixer is required for hand-truncated fragments (raw chain-cut PDB lacks proper terminal-cap templates that amber14 expects). Found-residue caching via findMissingAtoms + addMissingAtoms produces a clean N-terminal NH + C-terminal COO.
3. charge_from_molecules must go via template_generator_kwargs in openmmforcefields.SystemGenerator — not as a top-level kwarg. Pre-assigned partial charges on the openff Molecule are otherwise re-computed and the am1bcc path crashes.
4. RAMDSimulation requires properties dict when platform is set (e.g. {"Precision": "mixed"} for OpenCL).
5. sim.RAMD_step is a method, not an attribute — extract exit step from smoke_ramd.log regex ^EXIT:\s+(\d+) instead.
6. Performance on Apple Silicon Metal-OpenCL: ~50 RAMD-evaluations/min (≈ 5 ps biased-MD/min) for the 45.7 k-atom system. 17.9 ps unbinding ⇒ ~4 min biased-MD + 6 min build/equilibration overhead.
7. For 10 replicas × 14 ligands shortlist (assuming similar per-replica wall ~1–10 min): 1.5–25 hr total serial; 4× parallelism via independent processes brings it under one overnight.

Ranking delta caveats

• One replica is not statistically meaningful. Single τ_obs = 17.9 ps establishes ballpark + pipeline; production needs 10 replicas/ligand for SEM, with a ≥ 3-replica unbinding fraction reported.
• Phase-5d-truncated vs AF3-folded substrate: smoke used the former; Phase 5l selection gate identifies 151_e1659a AF3 model as the next-step substrate for the production run. A pre-flight comparison (1 replica each on both substrates) would tell us whether substrate choice biases τ ranking — recommended before committing the shortlist run.

Phase 5l final selection

Substrate for Phase 5m production: AF3 fragment 151_e1659a (fold_strc_k1141_frag_151_e1659a_model_0.cif, ring RMSD 0.97 Å vs Phase 5d, pocket pLDDT 78.7, pTM 0.65).

Hand-truncated Phase-5d fragment retained as secondary-substrate sensitivity check (1 replica vs AF3-folded; deviation > 30 % in τ would flag substrate choice as a confound).

Ranking delta

• tier A → A (held)
• mech 4 → 4 (held; pocket geometry preserved across AF3 fragment ladder, matches Phase 5d MD-relaxed parent at the K1141 ring level)
• deliv 3 → 3 (held; selectivity gate FAIL still binding constraint per Phase 8g-v3-lite)
• misha_fit 4 → 4 (held)
• next_step refresh: “Phase 5m production (10 replicas × 14 v5.2 shortlist + 1 replica diflunisal positive control + 1 replica tafamidis-analog negative control on 151_e1659a AF3 substrate, am1bcc charges if AmberTools installable, otherwise Gasteiger; expected wall ~5–25 hr depending on parallelism); cross-substrate sensitivity check (1 replica × 4 ligands on Phase-5d-truncated fragment) to validate substrate-independence; Boltz-2 binding-mode at TRPM4 + TMEM16A still queued P0-light; AF3 fragment ladder PASS (relaxed gate).”

Connections

• [depends-on] STRC h01 Phase 8g-v3-lite + 8h-lite 9 Pocket Max Phi + Dose Sweep 2026-04-26
• [implements] Pharmacochaperone Residence Time Criterion
• [uses] STRC K1141 Fragment Construct Strategy
• [uses] RAMD-OpenMM
• [informs] STRC E1659A QC Rate-Limiting Step Lit Gap-Map
• [part-of] H01 hub
• [about] Misha